Robust estimation of outage costs in South Korea using a machine learning technique: Bayesian Tobit quantile regression

As the industrial structure of the modern industry becomes more sophisticated and interdependent, accurate evaluation of customers' costs from power outages becomes increasingly difficult, but important. In this study, we propose a novel method to accurately evaluate customers' outage cost, Bayesian Tobit quantile regression. Using Bayesian Tobit quantile regression and survey data on customers' willingness to pay (WTP) to avoid power outages, we estimate customer damage functions (CDF) for the four industrial sectors in South Korea and compare the estimated CDFs with the estimates from a standard Tobit regression that many previous studies have used. Our empirical results reveal the two limitations of the previous analyses: CDFs estimated from the standard Tobit regression provide inaccurate cost estimates for prolonged-outages (longer than 5 h), and outliers in the survey make the estimates biased for short-duration outages (less than 3.5 h). Meanwhile, by providing five conditional quantile regression curve estimates (i.e., 10%, 25%, 50%, 75%, and 90%), the results from the Bayesian Tobit quantile regression facilitate the development of a robust and comprehensive interpretation of customers' outage costs. We also investigate the relationships between customers' outage cost and their idiosyncratic characteristics, employee size and electricity consumption. The employee size is positively related to WTP for outage-vulnerable customers except for less vulnerable customers in the industry sector, and electricity consumption is positively related to WTP only for such outage-vulnerable customers in all sectors. The rich background information about customers' outage costs provided by our study will help policymakers develop advanced electricity supply plans.